Ex-pro subtitler open-sources Subtitle Foundry, keeping ASR, translation and timing errors separately inspectable

mrtdlgc · reddit · 2026-11-03

A former professional translator/subtitler explains the core design of Subtitle Foundry (AGPL-3.0, on GitHub): keep transcription, translation and timing separately inspectable even with SOTA speech/language tools. A fluent final subtitle can hide where an error entered — if ASR drops a negation, the translation model may recover the intent or confidently translate the wrong statement, and the final file can't tell you which happened.

Three checks: compare raw ASR output against audio (keep it even if later stages improve display); judge translation against source audio/corrected transcript in dialogue context (fluency isn't evidence an ASR correction was justified); review cue boundaries, speaker changes, line breaks and duration by ear.

Timing evidence needs a label too: the default route uses ElevenLabs Scribe v2 native word timestamps; speech-activity detection locates speech, not words. The app offers exact-timing and editorial modes, both ending in video/waveform review with saved manual corrections. For evaluation, fix audio excerpts and references, record model/settings, and assess stages separately — aggregate word-error rates don't measure translation quality or readability. Testing so far: Turkish and Ukrainian to English with ElevenLabs + Codex; no published end-to-end benchmark.

Original post →

Ex-pro subtitler open-sources Subtitle Foundry, keeping ASR, translation and timing errors separately inspectable

More from coding & agent